What AI Reveals About Legislative Capacity
Six lessons from US state legislatures, and why keeping a “human in the loop” may be harder than it sounds
BY BEATRIZ REY
I spent the morning drinking tea (after almost 30 years of being a heavy coffee drinker, I have now switched to Turkish tea, which I highly recommend) and reading POPVOX Foundation’s latest report, State Legislative Capacity in the AI Era, which looks at how state legislatures in the United States are beginning to adapt to AI.
Six findings stayed with me as I started thinking about what they might mean beyond the US context.
To my surprise, the question “Does legislature X use AI?” is not very useful. I think this matters because we are used to assessing parliaments according to their level of AI adoption as organizations. But different people and offices inside the same legislature may be at completely different stages of adoption. The authors cite NCSL data showing that reported staff use rose from 20% in 2024 to 44% in 2025. Formal AI policies have not followed the same trend.
AI may be widely available, but it will only be useful if a legislature has the staff time, knowledge systems, data, governance structures, procurement capacity, and expertise needed to make good use of it.
A lot of what makes legislatures work exists only in the memories of experienced staff. The report calls this the “single-expert problem”: important institutional capabilities can effectively reside in one person. What I found interesting is the paradox this creates for AI: the legislatures that could benefit most from AI-assisted knowledge management may be those least prepared to use it (because their procedures, precedents, and tacit expertise have never been systematically captured).
One idea that really caught my attention as a political scientist is that legislative capacity can be elastic. Formal staff numbers, salaries, and budgets do not tell us everything an institution can actually mobilize. Caucus staff, interns, universities, civil society organizations, informal networks, and outside experts can all expand a legislature’s effective capacity. But they can also create dependencies, and that capacity may be uneven or disappear when those relationships change. So when we compare parliaments, maybe we should ask not only what resources they formally have, but what they can actually mobilize — and how much of that capacity would survive if particular people or relationships disappeared.
The report finds evidence that AI can increase legislative workload. Yes, you read that right. Members can generate more draft ideas; constituents can produce longer messages; advocates and lobbyists can generate materials more cheaply; and legislative staff increasingly receive AI-generated drafts and legal memos that may require major correction. Maybe the question is not just how many staff a legislature has, but whether its resources are enough for the volume and complexity of the work coming in.
Having a “human in the loop” is not enough. The report documents legislators treating AI-generated output as authoritative — “AI says” or “Grok says.” The authors argue that human review has to mean source-based verification, particularly in legislative and legal work. This is why tools such as Legible (adopted in Iowa) and Skywolf (adopted in Arizona) are so interesting: they constrain the model to a defined body of legislative material, making verification easier than with a general-purpose system.
The human in the loop problem
I want to think a little more about the “human in the loop” problem. The basic idea is that humans should verify and question what AI systems produce. But, as the report suggests, that is easier to say than to do.
The World Bank’s Tiago Peixoto has pointed out that these frameworks often assume capacities that many public institutions do not have. Political scientist Michael Hallsworth raises another problem: if professionals are constantly asked to supervise AI outputs, that supervision may itself become exhausting and may gradually weaken the skills they are supposed to be using to check the machine.
For legislatures, I think this opens up a bigger question about modernization. It is not enough to ask whether a human is technically still “in the loop.” We also need to ask what happens to human judgment when more and more of the work is mediated by AI.
I recently wrote about how AI could reshape legislatures as information-processing institutions. But legislatures have never depended only on moving information around. They also depend on Members and staff being able to interpret competing claims, tell plausible arguments from fabricated ones, and make judgments under uncertainty.
“But legislatures have never depended only on moving information around. They also depend on Members and staff being able to interpret competing claims, tell plausible arguments from fabricated ones, and make judgments under uncertainty.”
So for me, the practical question is not only how legislatures adopt AI. It is also how they preserve the conditions that allow people to keep exercising judgment well.
That means avoiding workflows in which staff become little more than validators of machine output. It means making sure expertise is still developed, rather than replaced by supervision. And it means being realistic about institutional capacity: a model of AI governance that requires time, staffing, and expertise that many legislatures do not have is not much of a model.
I’ll keep thinking through these questions as part of my research on AI and human judgment in legislative settings. But they are also questions legislative modernizers, parliamentary staff, and institutional leaders need to consider now. If you are working on similar issues, or dealing with these tensions inside your own institution, I’d be very interested in exchanging ideas.
Modern Parliament (“ModParl”) is a newsletter from POPVOX Foundation that provides insights into the evolution of legislative institutions worldwide. Learn more and subscribe at modparl.substack.com.
